AI Bookkeeper files bank lines with four layers: your rules, payee memory, a naive Bayes model, and a built-in keyword list. A new probe tested the last layer on 72 descriptors from businesses the earlier probes had not covered: general contractors, consultants who travel, nonprofits and churches, and event venues and caterers.
The result before any new rules: 33 filed correctly, none wrong, 39 left with no suggestion. Airlines, hotels and ad platforms were already covered. The gaps were construction software (Procore, Buildertrend), equipment rental, bonds and builders-risk insurance, and nonprofit payout platforms (Pushpay, Givelify, Tithely). An unfiled line goes to the review queue, so the cost of a miss is a click, not a wrong ledger entry.
Two of the expectations were mine to correct: Mailchimp and ShipStation already file to advertising and shipping, which is reasonable. Costco and Sam's Club were dropped because the same card is often used for personal shopping.
After adding rules the probe scores 72 of 72. That number is a ceiling, not an accuracy: the descriptors were written by an agent from public company names and the rules were tuned on them. The honest next step is still labelled real books. Details are in the AI Bookkeeper README under "Next".